Low Power Optimization for MPI Collective Operations

DVFS-available (Dynamic Voltage/Frequency Scaling) processors make it possible for a system to reduce the energy consumption by scaling down the frequency/voltage of the processors in high performance computing. For MPI collective operations, network communication time occupies the most of the whole time. Scaling down CPU voltage/frequency in non-critical path can effectively reduce energy consumption. This paper proposes Low-Power MPI_Gather algorithm (LPMG) and Low-Power MPI_Scatter algorithm (LPMS) and extend them to almost all the MPI collective operations. We evaluate the effectiveness of our low-power MPI collective operation algorithm using Intel MPI benchmark IMB on 128-processor cluster system connected by a 1000 Mbps Ethernet. Experimental results show that different MPI collective operations can achieve different energy saving. With 128 processes, average 45.9% and 55.7% energy savings can be reached through LPMG and LPMS, respectively. But MPI_Alltoall only gets 2.2% energy saving.